Markov models are extensively used for categorical sequence clustering and classification due to their inherent ability to capture complex chronological dependencies hidden in sequential data. Existing Markov models are based on an implicit assumption that the probability of the next state depends on the preceding context/pattern which is consist of consecutive states. This restriction hampers the models since some patterns, disrupted by noise, may be not frequent enough in a consecutive form, but frequent in a sparse form, which can not make use of the information hidden in the sequential data. A sparse pattern corresponds to a pattern in which one or some of the state(s) between the first and last one in the pattern is/are replaced by wildcard(s) that can be matched by a subset of values in the state set. In this paper, we propose a new model that generalizes the conventional Markov approach making it capable of dealing with the sparse pattern and handling the length of the sparse patterns adaptively, i.e. allowing variable length pattern with variable wildcards. The model, named Dynamic order Markov model (DOMM), allows deriving a new similarity measure between a sequence and a set of sequences/cluster. DOMM builds a sparse pattern from sub-frequent patterns that contain significant statistical information veiled by the noise. To implement DOMM, we propose a sparse pattern detector (SPD) based on the probability suffix tree (PST) capable of discovering both sparse and consecutive patterns, and then we develop a divisive clustering algorithm, named DMSC, for Dynamic order Markov model for categorical sequence clustering. Experimental results on real-world datasets demonstrate the promising performance of the proposed model.
Keywords: Categorical sequence clustering; Dynamic order Markov model; Pattern detection; Sparse pattern.
© The Author(s) 2021.